A denoising representation framework for underwater acoustic signal recognition

Area of Science:

• Signal Processing
• Acoustics
• Machine Learning

Background:

• Underwater acoustic signal recognition is challenged by significant noise interference.
• Existing denoising methods may not adequately preserve target signal features.

Purpose of the Study:

• To propose a practical method for denoising underwater acoustic signals for improved recognition.
• To enhance the feature representation and classification accuracy of acoustic signals.

Main Methods:

• Adaptive generation of multi-images from noise and target signals using correlation and "dropout".
• Utilizing a convolutional denoising autoencoder for parallel training and feature extraction.
• Employing weight fusion for initializing parallel random forest (RF) classifiers to boost accuracy.

Main Results:

• The proposed method demonstrates superior performance in feature denoising compared to other techniques.
• Enhanced classification accuracy was achieved in underwater acoustic scenes.
• The algorithm effectively suppresses noise interference while retaining crucial signal characteristics.

Conclusions:

• The developed denoising representation and recognition method offers a practical solution for underwater acoustic signal processing.
• The combination of adaptive multi-image generation, convolutional autoencoders, and RF weight fusion leads to significant improvements.
• This approach shows promise for robust target recognition in noisy marine environments.